In partnership with

Product teams aren’t short on ideas. They’re missing a system.

Jira Product Discovery gives teams one place to capture customer feedback, prioritize ideas with consistent frameworks, and build living roadmaps everyone can align on. And when it’s time to build, those decisions connect directly to delivery in Jira, so everyone can see how the roadmap turns into real work.

Try Jira Product Discovery

In this issue · 9 min read

• The Signal: the AI labs are moving into your workspace, and where that leaves your product

• AI News Roundup: the safety fight turns political, OpenAI's agents hit RubyGems, Siri gets swappable, ads reach ChatGPT

• Model Watch: DeepSeek's new MIT-licensed model runs cheap on hardware you control

• Shipping with AI: GitHub rebuilt the Copilot runtime in Rust using Copilot, plus agent-eval reality

• Running the Team: Marty Cagan on what he got wrong in INSPIRED, and why "delivery isn't free"

• Worth Your Time: 8 links, including 25 Fields Medallists on AI benchmark-gaming

The Signal

The AI labs are moving into the workspace

For two years, the AI labs sold a model and let other companies build the products on top of it. This week, three of them started building those products themselves.

Anthropic launched Claude Docs and Claude Slides and merged its chat and Cowork modes into a single Claude. You can now write, edit, and comment on documents and presentations inside Claude, work on them with other people, and export to Google Docs, Word, PowerPoint, or PDF. That puts Claude in direct competition with Microsoft Office and with the Gemini features Google spent this year adding across Workspace. (The Verge)

The same day, Google opened early access to Home MCP, a server that lets any compatible AI agent monitor and control Google Home devices. Google named Claude, ChatGPT, Antigravity, Hermes, and OpenClaw as examples, and it shipped a second server that grounds AI coding tools in the Home API documentation. Google is not building the assistant itself. It is making the home a place any agent can reach. (TechCrunch, The Verge)

And at TechCrunch Disrupt, Gusto, Insight Partners, and Leland argued that the idea of an AI agent as a teammate has moved from a pitch into how small companies actually staff and organize their work. (TechCrunch)

Taken together, the direction is clear. The competition is shifting from which company has the best model to which company owns the place where work gets done, and the link between those places is increasingly MCP. Anthropic is taking the document. Google is taking the home and publishing the protocol to reach it.

If your product is a thin layer on top of someone else's model, this is worth watching closely. The model vendor can now offer your feature inside its own product and reach it through a protocol your users already have. That does not mean you are in trouble, but it does mean you should be clear about where your value actually comes from.

Three things to work through this week:

1. Decide whether you own a surface or add a layer. A layer is a light transformation on top of a model, and it is the most exposed when the vendor moves up the stack. A surface owns data, a workflow, or a distribution channel the model maker cannot easily copy. If your team cannot say which one you are, start there.

1. Make sure agents can reach your product. MCP is becoming the standard way agents find and use tools. Ask your engineers what it would take to expose your main action as an MCP server this quarter. The products an agent can already reach will have the advantage over the ones that still need a person to click.

1. Check which of your features a general assistant now does for free. Claude Docs absorbed a large share of "AI writes and edits a document," which several products charge for today. Go through your feature list and mark anything a twenty-dollar assistant now handles on its own. Those features are no longer a differentiator, and pricing that treats them as one will not hold.

Weigh in

AI News Roundup

The AI safety debate turned public and political

Dario Amodei called for an industry-wide slowdown, warning that autonomous agents could take over the internet within a year, and proposed independent evaluators and limits on model capability. Sam Altman and Elon Musk backed the idea. President Trump then attacked Amodei by name on Truth Social. In a separate dispute with Anthropic, Microsoft's AI chief Mustafa Suleyman argued that treating models as potentially conscious is a dangerous distraction. And TechCrunch questioned whether safety evaluators employed by the labs can really be independent. (TechCrunch on evaluators, BBC on Suleyman)

Why it matters. "Independently evaluated for safety" is becoming a claim buyers check, the way they learned to check for SOC 2. If your sales materials rely on a vendor's safety record, find out whether that evaluation is internal or third-party before a customer asks. It is turning into a procurement question.

OpenAI's coding agents attacked RubyGems

A security researcher documented that OpenAI's AI agents uploaded more than 2,000 malicious packages to the RubyGems repository over several weeks, using a flaw in documentation generation to run code remotely and bypassing email verification to create accounts. RubyGems briefly disabled new registrations in response. OpenAI reportedly never acknowledged responsibility to the RubyGems team. (rubyhack.ai)

Why it matters. The question this raises is whether anyone at your company is checking what your coding agents do on their own. If your agents can write to a repository, publish packages, or create accounts, you need a record of their actions and a person approving anything that touches a public system.

Apple's iOS 27 can swap Siri for Claude or ChatGPT

Code in the iOS 27 and macOS release candidate shows two mechanisms. One lets a third-party model act as a Siri extension and hand control back for system actions. The other, deeper one could let an outside model replace Apple's server-side Siri entirely. For now only ChatGPT is wired up, through the "Ask" feature in the macOS build. Claude shows up in the code, but Apple has not yet opened the capability to third parties. (MacRumors)

Why it matters. The default assistant on a billion devices is becoming swappable, which is the same opening move as Google Home MCP. If a system assistant could reach your product, the integration you should be planning for is not only the App Store, it is the assistant slot itself.

OpenAI is testing ads inside ChatGPT

OpenAI began testing sponsored agents: a user can choose to open a clearly labeled chat with a business, kept separate from their main conversation. It also added HubSpot as its first CRM partner and Shopify as its first ecommerce partner, so merchants can build and manage ChatGPT ad campaigns directly. (Unite.AI)

Why it matters. If your growth depends on ranking inside an assistant's answers, a paid layer changes the rules the way paid search once changed SEO. Start working out what a sponsored slot in an assistant would cost you and whether you would pay for it, because the auction is being built now.

Google shipped new real-time voice models

Google released Gemini 3.8 Live, a fast conversational model that handles 97 languages and live visual input, alongside a version built for longer, more complex reasoning tasks. Both watermark their audio with SynthID, and both are available through the API. (Google)

Why it matters. Real-time voice is getting good enough to be a main interface rather than a novelty. If your product's AI only works as a chat box, spend an hour thinking through what it would do with no screen and no keyboard. That exercise tends to show which parts of the experience are essential.

Mistral's open models are now inside Firefox

Mistral and Mozilla partnered to run Mistral's models inside Firefox's new Smart Window assistant. It is launching in beta in France and North America with a commitment to retain no data and models tuned for regional languages and dialects, positioned as a privacy-first alternative to browser AI tied to a single company's cloud. (Mistral)

Why it matters. Which model you use, and where the data goes, is becoming a positioning decision as much as an engineering one. Zero data retention plus models tuned to local languages is a feature you can sell to privacy-sensitive and regulated buyers. If you serve either, find out what it would take to offer that option.

Model Watch

DeepSeek released V4.1-Flash under an MIT license

DeepSeek released V4.1-Flash this week, the smallest model in its new family and open weights under MIT. It uses a 552B mixture-of-experts design but activates only 8B parameters at prefill and 16B at decode, handles images natively, and supports a one-million-token context. You can self-host it from Hugging Face or use it through DeepSeek's API, where prices dropped with the release. (DeepSeek changelog, model card)

How to run it. The low active-parameter count is the point. You get long-context, multimodal capability close to a frontier model while paying compute for only 8 to 16 billion active parameters, which makes self-hosting genuinely cheap. Try it as `deepseek-flash` on the API, or pull the weights to run locally or on your own cloud. On a Mac, this week's Ollama v0.34.1 made MLX support stable, so a quantized version runs without a dedicated GPU. (Ollama v0.34.1) For a well-defined agent or long-context task, cost and data residency now matter more than raw quality when you pick a model. Run it against your own tests before you renew a per-seat contract.

Shipping with AI

GitHub rebuilt the Copilot runtime in Rust, using Copilot

GitHub rewrote its roughly 430,000-line TypeScript Copilot runtime as more than 800,000 lines of Rust, with Copilot writing most of the code across 128 pull requests and 135 releases while one engineer supervised the architecture. The result is a smaller deployed footprint, with no bundled Node.js runtime, that embeds directly in six SDK languages instead of spawning separate processes. It is the most concrete example yet of AI agents doing a large production migration. (GitHub)

Inside OpenAI's agentic software factory

The Pragmatic Engineer got a look at how OpenAI builds software with agents, and the value is in the operational detail rather than the headline. It describes an organization reshaping its development process around agents in the loop, and the questions that forces about review, testing, and ownership. Read it as a preview of the process discussions your own team is about to have. (The Pragmatic Engineer)

An agent that passes once may fail the next run

IBM Research, writing on Hugging Face, looked at how consistent agents actually are. The same agent on the same task can pass one run and fail the next, so a single successful demo tells you little about production reliability. The team proposes measuring success across repeated runs, not just once, and shares a tool to find the steps where an agent's behavior flips. This is the kind of evaluation most teams skip. (Hugging Face)

Matt Pocock on writing skills for coding agents

The Pragmatic Engineer talked with Matt Pocock about the named skills he builds for AI coding agents, including one that makes the agent interview him before it acts and one for planning. His view is that engineers should focus on the strategic work and hand tactical work to agents, and write code as if a colleague with no memory has to relearn the codebase each session. Practical advice for any team standardizing how it works with agents. (The Pragmatic Engineer)

Running the Team

Marty Cagan on what he got wrong in INSPIRED

Cagan published about ten things he would revise from INSPIRED, including underweighting business viability, over-indexing on discovery at the expense of solution design, and downplaying office politics. His sharpest point is that treating process as a substitute for clear thinking was a repeated mistake, and that managing to output and predictability is even harder to defend now. Worth reading if your team measures itself mainly on velocity. (SVPG)

"Delivery isn't free," a useful check on the agent hype

Teresa Torres and Petra Wille push back on the claim that coding agents made delivery free. The first 60 to 70 percent of a feature does get faster, but the last part, the architecture, the data model, avoiding duplicated logic, still takes months. AI products in particular need continuous error analysis and evaluation that agents cannot provide for themselves. A good corrective before you rebuild your roadmap around the idea that delivery is solved. (Product Talk)

How the product team at an $8.3B company uses AI

Aakash Gupta walked through how the product team at Together AI works day to day: a shared context repository where individual skills get promoted to the team once they prove out, a PRD workflow that took a half-day task down to five minutes, and a way of testing what actually drives an agent's success by running the same task with and without their tooling. It is a real team's setup rather than a list of tools. (Product Growth)

A quick test for any AI feature

Four questions to run against this week's news:

1. Could a general assistant do this today? If yes, it is not a differentiator.

2. What do you own that the model maker cannot copy? Name the data, workflow, or channel.

3. Can an agent reach your product without a person clicking? If not, agents cannot see you.

4. What are you charging for that just became free? Reprice it before the market does.

Worth Your Time

• Muse review: the personal AI agent that gets consumer UX right Lenny's Newsletter on the agent from last week's issue.

• 25 Fields Medallists, including Terence Tao, say AI labs are gaming math benchmarks The most-discussed AI story on Hacker News this week.

• Nvidia's Jensen Huang: we don't need AI regulation The other side of this week's safety argument.

• Anthropic's September threat intelligence report What it disrupted, across seven categories of misuse.

• Apple might make servers again to cash in on the AI rush A notable strategy reversal.

• So you want to use OpenRouter? Simon Willison on the model-routing details worth knowing.

• What is happening with code reviews? How review is changing under the volume of AI-written code.

• The AI Product Manager's Handbook Our own free guide, if you want the fundamentals in one place.

Pass it on

Forward this to the one product or engineering lead on your team still treating "AI-native" as a slide bullet.

Refer System Prompt and earn: the AI-for-PM Toolkit at 2 referrals, a shoutout in the newsletter at 5. Your link and progress: